Brain Tumor Recurrence vs. Radiation Necrosis Classification and Patient Survivability Prediction.

Area of Science:

• Radiology
• Medical Imaging
• Computational Biology

Background:

• Glioblastoma (GB) is an aggressive brain tumor with poor prognosis.
• Post-treatment MRI changes can indicate either radiation necrosis (RN) or recurrent brain tumor (rBT), posing a diagnostic challenge.
• Early differentiation of rBT from RN is critical for timely patient management and improved outcomes.

Purpose of the Study:

• To develop and validate a computational model for differentiating recurrent brain tumors from radiation necrosis using MRI.
• To assess the utility of multiresolution radiomic features (MRF) for classification and survival prediction in GB patients.
• To address potential class imbalance issues in machine learning models for rBT/RN classification.

Main Methods:

• Extraction of multiresolution radiomic features (MRF) from MRI scans.
• Application of statistical significance testing (p<0.05) for feature selection.
• Utilized repeated random sub-sampling to balance dataset for classification tasks.
• Implemented five-fold cross-validation for model evaluation.

Main Results:

• The MRF-based model achieved an AUC of 0.892±0.055 for classifying recurrent brain tumors from radiation necrosis.
• The model demonstrated feasibility as a non-invasive biomarker for predicting patient risk of recurrence or necrosis.
• Overall survival prediction yielded an AUC of 0.77±0.032, outperforming comparisons.

Conclusions:

• The proposed computational model effectively distinguishes between recurrent brain tumors and radiation necrosis.
• Multiresolution radiomic features serve as a promising non-invasive biomarker for patient stratification and survival prediction.
• The method offers a statistically rigorous approach to handle class imbalance in medical image analysis.